kernel - lwkt_token revamp
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
3b998fa9 2 * Copyright (c) 2003-2010 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
47#include <sys/queue.h>
7d0bac62 48#include <sys/sysctl.h>
99df837e 49#include <sys/kthread.h>
f1d1c3fa 50#include <machine/cpu.h>
99df837e 51#include <sys/lock.h>
f6bf3af1 52#include <sys/caps.h>
9d265729 53#include <sys/spinlock.h>
57aa743c 54#include <sys/ktr.h>
9d265729
MD
55
56#include <sys/thread2.h>
57#include <sys/spinlock2.h>
684a93c4 58#include <sys/mplock2.h>
f1d1c3fa 59
8c72e3d5
AH
60#include <sys/dsched.h>
61
7d0bac62
MD
62#include <vm/vm.h>
63#include <vm/vm_param.h>
64#include <vm/vm_kern.h>
65#include <vm/vm_object.h>
66#include <vm/vm_page.h>
67#include <vm/vm_map.h>
68#include <vm/vm_pager.h>
69#include <vm/vm_extern.h>
7d0bac62 70
99df837e 71#include <machine/stdarg.h>
96728c05 72#include <machine/smp.h>
99df837e 73
d850923c
AE
74#if !defined(KTR_CTXSW)
75#define KTR_CTXSW KTR_ALL
76#endif
77KTR_INFO_MASTER(ctxsw);
a1f0fb66
AE
78KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
79 sizeof(int) + sizeof(struct thread *));
80KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
81 sizeof(int) + sizeof(struct thread *));
82KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
83 sizeof (struct thread *) + sizeof(char *));
84KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
1541028a 85
40aaf5fc
NT
86static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
87
0f7a3396
MD
88#ifdef INVARIANTS
89static int panic_on_cscount = 0;
90#endif
05220613
MD
91static __int64_t switch_count = 0;
92static __int64_t preempt_hit = 0;
93static __int64_t preempt_miss = 0;
94static __int64_t preempt_weird = 0;
f64b567c 95static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 96static int lwkt_use_spin_port;
40aaf5fc 97static struct objcache *thread_cache;
05220613 98
88ebb169 99#ifdef SMP
e381e77c 100static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 101#endif
e381e77c 102
0855a2af
JG
103extern void cpu_heavy_restore(void);
104extern void cpu_lwkt_restore(void);
105extern void cpu_kthread_restore(void);
106extern void cpu_idle_restore(void);
107
b2b3ffcd 108#ifdef __x86_64__
85514115
MD
109
110static int
0855a2af
JG
111jg_tos_ok(struct thread *td)
112{
85514115
MD
113 void *tos;
114 int tos_ok;
115
0855a2af
JG
116 if (td == NULL) {
117 return 1;
118 }
119 KKASSERT(td->td_sp != NULL);
85514115
MD
120 tos = ((void **)td->td_sp)[0];
121 tos_ok = 0;
122 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
123 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
124 tos_ok = 1;
125 }
126 return tos_ok;
127}
128
85514115
MD
129#endif
130
fb0f29c4
MD
131/*
132 * We can make all thread ports use the spin backend instead of the thread
133 * backend. This should only be set to debug the spin backend.
134 */
135TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
136
0f7a3396
MD
137#ifdef INVARIANTS
138SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
139#endif
4b5f931b 140SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 141SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 142SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 143SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
144#ifdef INVARIANTS
145SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
146 &token_contention_count, 0, "spinning due to token contention");
38717797 147#endif
05220613 148
4b5f931b
MD
149/*
150 * These helper procedures handle the runq, they can only be called from
151 * within a critical section.
75cdbe6c
MD
152 *
153 * WARNING! Prior to SMP being brought up it is possible to enqueue and
154 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
155 * instead of 'mycpu' when referencing the globaldata structure. Once
156 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 157 */
f1d1c3fa
MD
158static __inline
159void
160_lwkt_dequeue(thread_t td)
161{
162 if (td->td_flags & TDF_RUNQ) {
4b5f931b 163 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 164 struct globaldata *gd = td->td_gd;
4b5f931b 165
f1d1c3fa 166 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
167 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
168 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
169 }
170}
171
172static __inline
173void
174_lwkt_enqueue(thread_t td)
175{
7f5d7ed7 176 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
4b5f931b 177 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 178 struct globaldata *gd = td->td_gd;
4b5f931b 179
f1d1c3fa 180 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
181 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
182 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
183 }
184}
8ad65e08 185
40aaf5fc
NT
186static __boolean_t
187_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
188{
189 struct thread *td = (struct thread *)obj;
190
191 td->td_kstack = NULL;
192 td->td_kstack_size = 0;
193 td->td_flags = TDF_ALLOCATED_THREAD;
194 return (1);
195}
196
197static void
198_lwkt_thread_dtor(void *obj, void *privdata)
199{
200 struct thread *td = (struct thread *)obj;
201
202 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
203 ("_lwkt_thread_dtor: not allocated from objcache"));
204 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
205 td->td_kstack_size > 0,
206 ("_lwkt_thread_dtor: corrupted stack"));
207 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
208}
209
210/*
211 * Initialize the lwkt s/system.
212 */
213void
214lwkt_init(void)
215{
216 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
217 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
218 NULL, CACHE_NTHREADS/2,
219 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
220}
221
37af14fe
MD
222/*
223 * Schedule a thread to run. As the current thread we can always safely
224 * schedule ourselves, and a shortcut procedure is provided for that
225 * function.
226 *
227 * (non-blocking, self contained on a per cpu basis)
228 */
229void
230lwkt_schedule_self(thread_t td)
231{
232 crit_enter_quick(td);
37af14fe 233 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 234 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 235 _lwkt_enqueue(td);
37af14fe
MD
236 crit_exit_quick(td);
237}
238
239/*
240 * Deschedule a thread.
241 *
242 * (non-blocking, self contained on a per cpu basis)
243 */
244void
245lwkt_deschedule_self(thread_t td)
246{
247 crit_enter_quick(td);
37af14fe
MD
248 _lwkt_dequeue(td);
249 crit_exit_quick(td);
250}
251
8ad65e08
MD
252/*
253 * LWKTs operate on a per-cpu basis
254 *
73e4f7b9 255 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
256 */
257void
258lwkt_gdinit(struct globaldata *gd)
259{
4b5f931b
MD
260 int i;
261
262 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
263 TAILQ_INIT(&gd->gd_tdrunq[i]);
264 gd->gd_runqmask = 0;
73e4f7b9 265 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
266}
267
268/*
7d0bac62 269 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
270 * or LWKT start address before it can be scheduled. If the target cpu is
271 * -1 the thread will be created on the current cpu.
0cfcada1
MD
272 *
273 * If you intend to create a thread without a process context this function
274 * does everything except load the startup and switcher function.
7d0bac62
MD
275 */
276thread_t
d3d32139 277lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 278{
c070746a 279 globaldata_t gd = mycpu;
99df837e 280 void *stack;
7d0bac62 281
c070746a
MD
282 /*
283 * If static thread storage is not supplied allocate a thread. Reuse
284 * a cached free thread if possible. gd_freetd is used to keep an exiting
285 * thread intact through the exit.
286 */
ef0fdad1 287 if (td == NULL) {
c070746a
MD
288 if ((td = gd->gd_freetd) != NULL)
289 gd->gd_freetd = NULL;
290 else
291 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
292 KASSERT((td->td_flags &
293 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
294 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
295 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 296 }
c070746a
MD
297
298 /*
299 * Try to reuse cached stack.
300 */
f470d0c8
MD
301 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
302 if (flags & TDF_ALLOCATED_STACK) {
e4846942 303 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
304 stack = NULL;
305 }
306 }
307 if (stack == NULL) {
e4846942 308 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 309 flags |= TDF_ALLOCATED_STACK;
99df837e 310 }
75cdbe6c 311 if (cpu < 0)
c070746a 312 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 313 else
f470d0c8 314 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 315 return(td);
7d0bac62
MD
316}
317
318/*
319 * Initialize a preexisting thread structure. This function is used by
320 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
321 *
f8c3996b
MD
322 * All threads start out in a critical section at a priority of
323 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
324 * appropriate. This function may send an IPI message when the
325 * requested cpu is not the current cpu and consequently gd_tdallq may
326 * not be initialized synchronously from the point of view of the originating
327 * cpu.
328 *
329 * NOTE! we have to be careful in regards to creating threads for other cpus
330 * if SMP has not yet been activated.
7d0bac62 331 */
41a01a4d
MD
332#ifdef SMP
333
75cdbe6c
MD
334static void
335lwkt_init_thread_remote(void *arg)
336{
337 thread_t td = arg;
338
52eedfb5
MD
339 /*
340 * Protected by critical section held by IPI dispatch
341 */
75cdbe6c
MD
342 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
343}
344
41a01a4d
MD
345#endif
346
7d0bac62 347void
f470d0c8
MD
348lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
349 struct globaldata *gd)
7d0bac62 350{
37af14fe
MD
351 globaldata_t mygd = mycpu;
352
99df837e
MD
353 bzero(td, sizeof(struct thread));
354 td->td_kstack = stack;
f470d0c8 355 td->td_kstack_size = stksize;
d3d32139 356 td->td_flags = flags;
26a0694b 357 td->td_gd = gd;
f8c3996b 358 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
3b998fa9 359 td->td_toks_stop = &td->td_toks_base;
d3d32139
MD
360#ifdef SMP
361 if ((flags & TDF_MPSAFE) == 0)
362 td->td_mpcount = 1;
363#endif
fb0f29c4
MD
364 if (lwkt_use_spin_port)
365 lwkt_initport_spin(&td->td_msgport);
366 else
367 lwkt_initport_thread(&td->td_msgport, td);
99df837e 368 pmap_init_thread(td);
0f7a3396 369#ifdef SMP
5d21b981
MD
370 /*
371 * Normally initializing a thread for a remote cpu requires sending an
372 * IPI. However, the idlethread is setup before the other cpus are
373 * activated so we have to treat it as a special case. XXX manipulation
374 * of gd_tdallq requires the BGL.
375 */
376 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 377 crit_enter_gd(mygd);
75cdbe6c 378 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 379 crit_exit_gd(mygd);
75cdbe6c 380 } else {
2db3b277 381 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 382 }
0f7a3396 383#else
37af14fe 384 crit_enter_gd(mygd);
0f7a3396 385 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 386 crit_exit_gd(mygd);
0f7a3396 387#endif
8c72e3d5
AH
388
389 dsched_new_thread(td);
73e4f7b9
MD
390}
391
392void
393lwkt_set_comm(thread_t td, const char *ctl, ...)
394{
e2565a42 395 __va_list va;
73e4f7b9 396
e2565a42 397 __va_start(va, ctl);
379210cb 398 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 399 __va_end(va);
e7c0dbba 400 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
401}
402
99df837e 403void
73e4f7b9 404lwkt_hold(thread_t td)
99df837e 405{
73e4f7b9
MD
406 ++td->td_refs;
407}
408
409void
410lwkt_rele(thread_t td)
411{
412 KKASSERT(td->td_refs > 0);
413 --td->td_refs;
414}
415
416void
417lwkt_wait_free(thread_t td)
418{
419 while (td->td_refs)
377d4740 420 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
421}
422
423void
424lwkt_free_thread(thread_t td)
425{
d9eea1a5 426 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
427 ("lwkt_free_thread: did not exit! %p", td));
428
40aaf5fc
NT
429 if (td->td_flags & TDF_ALLOCATED_THREAD) {
430 objcache_put(thread_cache, td);
431 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
432 /* client-allocated struct with internally allocated stack */
433 KASSERT(td->td_kstack && td->td_kstack_size > 0,
434 ("lwkt_free_thread: corrupted stack"));
435 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
436 td->td_kstack = NULL;
437 td->td_kstack_size = 0;
99df837e 438 }
e7c0dbba 439 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
440}
441
442
7d0bac62 443/*
8ad65e08 444 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
445 * switch to the idlethread. Switching must occur within a critical
446 * section to avoid races with the scheduling queue.
447 *
448 * We always have full control over our cpu's run queue. Other cpus
449 * that wish to manipulate our queue must use the cpu_*msg() calls to
450 * talk to our cpu, so a critical section is all that is needed and
451 * the result is very, very fast thread switching.
452 *
96728c05
MD
453 * The LWKT scheduler uses a fixed priority model and round-robins at
454 * each priority level. User process scheduling is a totally
455 * different beast and LWKT priorities should not be confused with
456 * user process priorities.
f1d1c3fa 457 *
96728c05
MD
458 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
459 * cleans it up. Note that the td_switch() function cannot do anything that
460 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
461 * the target thread (not the current thread). It's nice to have a scheduler
462 * that does not need the MP lock to work because it allows us to do some
463 * really cool high-performance MP lock optimizations.
69d78e99
MD
464 *
465 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
466 * is not called by the current thread in the preemption case, only when
467 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
468 */
469void
470lwkt_switch(void)
471{
37af14fe
MD
472 globaldata_t gd = mycpu;
473 thread_t td = gd->gd_curthread;
8ad65e08 474 thread_t ntd;
8a8d5d85
MD
475#ifdef SMP
476 int mpheld;
477#endif
8ad65e08 478
46a3f46d 479 /*
27e88a6e
MD
480 * Switching from within a 'fast' (non thread switched) interrupt or IPI
481 * is illegal. However, we may have to do it anyway if we hit a fatal
482 * kernel trap or we have paniced.
483 *
484 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 485 */
27e88a6e
MD
486 if (gd->gd_intr_nesting_level) {
487 int savegdnest;
488 int savegdtrap;
489
490 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
491 panic("lwkt_switch: cannot switch from within "
492 "a fast interrupt, yet, td %p\n", td);
493 } else {
494 savegdnest = gd->gd_intr_nesting_level;
495 savegdtrap = gd->gd_trap_nesting_level;
496 gd->gd_intr_nesting_level = 0;
497 gd->gd_trap_nesting_level = 0;
a7422615
MD
498 if ((td->td_flags & TDF_PANICWARN) == 0) {
499 td->td_flags |= TDF_PANICWARN;
6ea70f76 500 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 501 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 502 print_backtrace(-1);
a7422615 503 }
27e88a6e
MD
504 lwkt_switch();
505 gd->gd_intr_nesting_level = savegdnest;
506 gd->gd_trap_nesting_level = savegdtrap;
507 return;
508 }
96728c05 509 }
ef0fdad1 510
cb973d15
MD
511 /*
512 * Passive release (used to transition from user to kernel mode
513 * when we block or switch rather then when we enter the kernel).
514 * This function is NOT called if we are switching into a preemption
515 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
516 * our current process designation (if we have one) and become a true
517 * LWKT thread, and may also hand the current process designation to
518 * another process and schedule thread.
cb973d15
MD
519 */
520 if (td->td_release)
521 td->td_release(td);
522
37af14fe 523 crit_enter_gd(gd);
3b998fa9 524 if (TD_TOKS_HELD(td))
9d265729
MD
525 lwkt_relalltokens(td);
526
527 /*
b02926de
MD
528 * We had better not be holding any spin locks, but don't get into an
529 * endless panic loop.
9d265729 530 */
bbb31c5d
MD
531 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
532 ("lwkt_switch: still holding a shared spinlock %p!",
533 gd->gd_spinlock_rd));
d666840a
MD
534 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
535 ("lwkt_switch: still holding %d exclusive spinlocks!",
536 gd->gd_spinlocks_wr));
9d265729 537
8a8d5d85
MD
538
539#ifdef SMP
540 /*
541 * td_mpcount cannot be used to determine if we currently hold the
542 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
543 * to get the lock, and switch out if it can't. Our ownership of
544 * the actual lock will remain stable while we are in a critical section
545 * (but, of course, another cpu may own or release the lock so the
546 * actual value of mp_lock is not stable).
8a8d5d85
MD
547 */
548 mpheld = MP_LOCK_HELD();
0f7a3396
MD
549#ifdef INVARIANTS
550 if (td->td_cscount) {
6ea70f76 551 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
552 td);
553 if (panic_on_cscount)
554 panic("switching while mastering cpusync");
555 }
556#endif
8a8d5d85 557#endif
99df837e
MD
558 if ((ntd = td->td_preempted) != NULL) {
559 /*
560 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
561 * thread. This occurs transparently, whether the preempted thread
562 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
563 * itself).
564 *
565 * We have to setup the MP lock for the original thread after backing
566 * out the adjustment that was made to curthread when the original
567 * was preempted.
99df837e 568 */
26a0694b 569 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 570#ifdef SMP
96728c05 571 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 572 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
573 td, ntd, td->td_mpcount, ntd->td_mpcount);
574 }
8a8d5d85
MD
575 if (ntd->td_mpcount) {
576 td->td_mpcount -= ntd->td_mpcount;
577 KKASSERT(td->td_mpcount >= 0);
578 }
579#endif
26a0694b 580 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
581
582 /*
b9eb1c19
MD
583 * The interrupt may have woken a thread up, we need to properly
584 * set the reschedule flag if the originally interrupted thread is
585 * at a lower priority.
8ec60c3f
MD
586 */
587 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
588 need_lwkt_resched();
8a8d5d85 589 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 590 } else {
4b5f931b
MD
591 /*
592 * Priority queue / round-robin at each priority. Note that user
593 * processes run at a fixed, low priority and the user process
594 * scheduler deals with interactions between user processes
595 * by scheduling and descheduling them from the LWKT queue as
596 * necessary.
8a8d5d85
MD
597 *
598 * We have to adjust the MP lock for the target thread. If we
599 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
600 * thread that does not need the MP lock. If we cannot, we spin
601 * instead of HLT.
602 *
603 * A similar issue exists for the tokens held by the target thread.
604 * If we cannot obtain ownership of the tokens we cannot immediately
605 * schedule the thread.
606 */
607
608 /*
8ec60c3f
MD
609 * If an LWKT reschedule was requested, well that is what we are
610 * doing now so clear it.
611 */
612 clear_lwkt_resched();
4b5f931b
MD
613again:
614 if (gd->gd_runqmask) {
615 int nq = bsrl(gd->gd_runqmask);
616 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
617 gd->gd_runqmask &= ~(1 << nq);
618 goto again;
619 }
8a8d5d85 620#ifdef SMP
41a01a4d 621 /*
df6b8ba0
MD
622 * THREAD SELECTION FOR AN SMP MACHINE BUILD
623 *
41a01a4d
MD
624 * If the target needs the MP lock and we couldn't get it,
625 * or if the target is holding tokens and we could not
626 * gain ownership of the tokens, continue looking for a
627 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
628 *
629 * NOTE: the mpheld variable invalid after this conditional, it
630 * can change due to both cpu_try_mplock() returning success
9d265729 631 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
632 * we are trying to check the mpcount of a thread other then
633 * the current thread. Because of this, if the current thread
634 * is not holding td_mpcount, an IPI indirectly run via
9d265729 635 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 636 * cause the core MP lock to be released.
41a01a4d
MD
637 */
638 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
3b998fa9 639 (TD_TOKS_HELD(ntd) && lwkt_getalltokens(ntd) == 0)
41a01a4d 640 ) {
8a8d5d85 641 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
642
643 mpheld = MP_LOCK_HELD();
644 ntd = NULL;
8a8d5d85
MD
645 while (rqmask) {
646 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 647 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 648 /* spinning due to MP lock being held */
41a01a4d 649 continue;
38717797 650 }
a453459d
MD
651
652 /*
9d265729 653 * mpheld state invalid after getalltokens call returns
a453459d
MD
654 * failure, but the variable is only needed for
655 * the loop.
656 */
3b998fa9 657 if (TD_TOKS_HELD(ntd) && !lwkt_getalltokens(ntd)) {
a453459d 658 /* spinning due to token contention */
38717797 659#ifdef INVARIANTS
a453459d 660 ++token_contention_count;
38717797 661#endif
a453459d 662 mpheld = MP_LOCK_HELD();
41a01a4d 663 continue;
38717797 664 }
41a01a4d 665 break;
8a8d5d85
MD
666 }
667 if (ntd)
668 break;
669 rqmask &= ~(1 << nq);
670 nq = bsrl(rqmask);
b9eb1c19
MD
671
672 /*
673 * We have two choices. We can either refuse to run a
674 * user thread when a kernel thread needs the MP lock
675 * but could not get it, or we can allow it to run but
676 * then expect an IPI (hopefully) later on to force a
677 * reschedule when the MP lock might become available.
678 */
679 if (nq < TDPRI_KERN_LPSCHED) {
684a93c4
MD
680 break; /* for now refuse to run */
681#if 0
b9eb1c19
MD
682 if (chain_mplock == 0)
683 break;
b9eb1c19 684 /* continue loop, allow user threads to be scheduled */
684a93c4 685#endif
b9eb1c19 686 }
8a8d5d85 687 }
684a93c4
MD
688
689 /*
690 * Case where a (kernel) thread needed the MP lock and could
691 * not get one, and we may or may not have found another
692 * thread which does not need the MP lock to run while
693 * we wait (ntd).
694 */
8a8d5d85 695 if (ntd == NULL) {
a2a5ad0d
MD
696 ntd = &gd->gd_idlethread;
697 ntd->td_flags |= TDF_IDLE_NOHLT;
684a93c4
MD
698 set_mplock_contention_mask(gd);
699 cpu_mplock_contested();
df6b8ba0 700 goto using_idle_thread;
8a8d5d85 701 } else {
684a93c4 702 clr_mplock_contention_mask(gd);
344ad853 703 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
704 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
705 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
706 }
707 } else {
684a93c4 708 clr_mplock_contention_mask(gd);
344ad853 709 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
710 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
711 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
712 }
713#else
df6b8ba0
MD
714 /*
715 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
716 * worry about tokens or the BGL. However, we still have
717 * to call lwkt_getalltokens() in order to properly detect
718 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 719 */
7eb611ef 720 lwkt_getalltokens(ntd);
344ad853 721 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
722 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
723 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 724#endif
4b5f931b 725 } else {
3c23a41a 726 /*
60f945af
MD
727 * We have nothing to run but only let the idle loop halt
728 * the cpu if there are no pending interrupts.
3c23a41a 729 */
a2a5ad0d 730 ntd = &gd->gd_idlethread;
60f945af 731 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 732 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 733#ifdef SMP
df6b8ba0
MD
734using_idle_thread:
735 /*
736 * The idle thread should not be holding the MP lock unless we
737 * are trapping in the kernel or in a panic. Since we select the
738 * idle thread unconditionally when no other thread is available,
739 * if the MP lock is desired during a panic or kernel trap, we
740 * have to loop in the scheduler until we get it.
741 */
742 if (ntd->td_mpcount) {
743 mpheld = MP_LOCK_HELD();
684a93c4 744 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
df6b8ba0 745 panic("Idle thread %p was holding the BGL!", ntd);
684a93c4 746 if (mpheld == 0)
df6b8ba0
MD
747 goto again;
748 }
a453459d 749#endif
4b5f931b 750 }
f1d1c3fa 751 }
26a0694b
MD
752 KASSERT(ntd->td_pri >= TDPRI_CRIT,
753 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
754
755 /*
756 * Do the actual switch. If the new target does not need the MP lock
757 * and we are holding it, release the MP lock. If the new target requires
758 * the MP lock we have already acquired it for the target.
759 */
760#ifdef SMP
761 if (ntd->td_mpcount == 0 ) {
762 if (MP_LOCK_HELD())
763 cpu_rel_mplock();
764 } else {
a453459d 765 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
766 }
767#endif
94f6d86e
MD
768 if (td != ntd) {
769 ++switch_count;
b2b3ffcd 770#ifdef __x86_64__
f64b567c
SW
771 {
772 int tos_ok __debugvar = jg_tos_ok(ntd);
773 KKASSERT(tos_ok);
774 }
85514115 775#endif
a1f0fb66 776 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 777 td->td_switch(ntd);
94f6d86e 778 }
37af14fe
MD
779 /* NOTE: current cpu may have changed after switch */
780 crit_exit_quick(td);
8ad65e08
MD
781}
782
f1d1c3fa 783/*
96728c05
MD
784 * Request that the target thread preempt the current thread. Preemption
785 * only works under a specific set of conditions:
b68b7282 786 *
96728c05
MD
787 * - We are not preempting ourselves
788 * - The target thread is owned by the current cpu
789 * - We are not currently being preempted
790 * - The target is not currently being preempted
d3d1cbc8
MD
791 * - We are not holding any spin locks
792 * - The target thread is not holding any tokens
96728c05
MD
793 * - We are able to satisfy the target's MP lock requirements (if any).
794 *
795 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
796 * this is called via lwkt_schedule() through the td_preemptable callback.
797 * critpri is the managed critical priority that we should ignore in order
798 * to determine whether preemption is possible (aka usually just the crit
799 * priority of lwkt_schedule() itself).
b68b7282 800 *
26a0694b
MD
801 * XXX at the moment we run the target thread in a critical section during
802 * the preemption in order to prevent the target from taking interrupts
803 * that *WE* can't. Preemption is strictly limited to interrupt threads
804 * and interrupt-like threads, outside of a critical section, and the
805 * preempted source thread will be resumed the instant the target blocks
806 * whether or not the source is scheduled (i.e. preemption is supposed to
807 * be as transparent as possible).
4b5f931b 808 *
8a8d5d85
MD
809 * The target thread inherits our MP count (added to its own) for the
810 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
811 * MP lock during the preemption. Therefore, any preempting targets must be
812 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
813 * out of sync with the physical mp_lock, but we do not have to preserve
814 * the original ownership of the lock if it was out of synch (that is, we
815 * can leave it synchronized on return).
b68b7282
MD
816 */
817void
96728c05 818lwkt_preempt(thread_t ntd, int critpri)
b68b7282 819{
46a3f46d 820 struct globaldata *gd = mycpu;
0a3f9b47 821 thread_t td;
8a8d5d85
MD
822#ifdef SMP
823 int mpheld;
57c254db 824 int savecnt;
8a8d5d85 825#endif
b68b7282 826
26a0694b 827 /*
96728c05
MD
828 * The caller has put us in a critical section. We can only preempt
829 * if the caller of the caller was not in a critical section (basically
d666840a 830 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 831 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 832 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
833 *
834 * YYY The target thread must be in a critical section (else it must
835 * inherit our critical section? I dunno yet).
41a01a4d 836 *
0a3f9b47 837 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
838 */
839 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 840
0a3f9b47 841 td = gd->gd_curthread;
0a3f9b47 842 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
843 ++preempt_miss;
844 return;
845 }
96728c05
MD
846 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
847 ++preempt_miss;
8ec60c3f 848 need_lwkt_resched();
96728c05
MD
849 return;
850 }
851#ifdef SMP
46a3f46d 852 if (ntd->td_gd != gd) {
96728c05 853 ++preempt_miss;
8ec60c3f 854 need_lwkt_resched();
96728c05
MD
855 return;
856 }
857#endif
41a01a4d 858 /*
d3d1cbc8 859 * Take the easy way out and do not preempt if we are holding
d666840a 860 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
861 * preempted interlock against the target thread's tokens and whether
862 * we can get all the target thread's tokens, but this situation
863 * should not occur very often so its easier to simply not preempt.
d666840a
MD
864 * Also, plain spinlocks are impossible to figure out at this point so
865 * just don't preempt.
d3d1cbc8
MD
866 *
867 * Do not try to preempt if the target thread is holding any tokens.
868 * We could try to acquire the tokens but this case is so rare there
869 * is no need to support it.
41a01a4d 870 */
bbb31c5d 871 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 872 ++preempt_miss;
8ec60c3f 873 need_lwkt_resched();
41a01a4d
MD
874 return;
875 }
3b998fa9 876 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
877 ++preempt_miss;
878 need_lwkt_resched();
879 return;
880 }
26a0694b
MD
881 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
882 ++preempt_weird;
8ec60c3f 883 need_lwkt_resched();
26a0694b
MD
884 return;
885 }
886 if (ntd->td_preempted) {
4b5f931b 887 ++preempt_hit;
8ec60c3f 888 need_lwkt_resched();
26a0694b 889 return;
b68b7282 890 }
8a8d5d85 891#ifdef SMP
a2a5ad0d
MD
892 /*
893 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
894 * to or from the MP lock. In this case td_mpcount will be pre-disposed
895 * (non-zero) but not actually synchronized with the actual state of the
896 * lock. We can use it to imply an MP lock requirement for the
897 * preemption but we cannot use it to test whether we hold the MP lock
898 * or not.
a2a5ad0d 899 */
96728c05 900 savecnt = td->td_mpcount;
71ef2f5c 901 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
902 ntd->td_mpcount += td->td_mpcount;
903 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
904 ntd->td_mpcount -= td->td_mpcount;
905 ++preempt_miss;
8ec60c3f 906 need_lwkt_resched();
8a8d5d85
MD
907 return;
908 }
909#endif
26a0694b 910
8ec60c3f
MD
911 /*
912 * Since we are able to preempt the current thread, there is no need to
913 * call need_lwkt_resched().
914 */
26a0694b
MD
915 ++preempt_hit;
916 ntd->td_preempted = td;
917 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 918 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
26a0694b 919 td->td_switch(ntd);
b9eb1c19 920
26a0694b 921 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
922#ifdef SMP
923 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
924 mpheld = MP_LOCK_HELD();
925 if (mpheld && td->td_mpcount == 0)
96728c05 926 cpu_rel_mplock();
71ef2f5c 927 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
928 panic("lwkt_preempt(): MP lock was not held through");
929#endif
26a0694b
MD
930 ntd->td_preempted = NULL;
931 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
932}
933
934/*
faaeffac 935 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 936 *
faaeffac
MD
937 * td_nest_count prevents deep nesting via splz() or doreti() which
938 * might otherwise blow out the kernel stack. Note that except for
939 * this special case, we MUST call splz() here to handle any
940 * pending ints, particularly after we switch, or we might accidently
941 * halt the cpu with interrupts pending.
ef0fdad1 942 *
f1d1c3fa
MD
943 * (self contained on a per cpu basis)
944 */
945void
faaeffac 946splz_check(void)
f1d1c3fa 947{
7966cb69
MD
948 globaldata_t gd = mycpu;
949 thread_t td = gd->gd_curthread;
ef0fdad1 950
46a3f46d 951 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 952 splz();
f1d1c3fa
MD
953}
954
8ad65e08 955/*
faaeffac
MD
956 * This implements a normal yield which will yield to equal priority
957 * threads as well as higher priority threads. Note that gd_reqflags
958 * tests will be handled by the crit_exit() call in lwkt_switch().
f1d1c3fa
MD
959 *
960 * (self contained on a per cpu basis)
8ad65e08
MD
961 */
962void
f1d1c3fa 963lwkt_yield(void)
8ad65e08 964{
37af14fe 965 lwkt_schedule_self(curthread);
f1d1c3fa
MD
966 lwkt_switch();
967}
968
969/*
3824f392
MD
970 * This function is used along with the lwkt_passive_recover() inline
971 * by the trap code to negotiate a passive release of the current
972 * process/lwp designation with the user scheduler.
973 */
974void
975lwkt_passive_release(struct thread *td)
976{
977 struct lwp *lp = td->td_lwp;
978
979 td->td_release = NULL;
980 lwkt_setpri_self(TDPRI_KERN_USER);
981 lp->lwp_proc->p_usched->release_curproc(lp);
982}
983
984/*
985 * Make a kernel thread act as if it were in user mode with regards
986 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
987 * loops which may be potentially cpu-bound can call lwkt_user_yield().
988 *
989 * The lwkt_user_yield() function is designed to have very low overhead
990 * if no yield is determined to be needed.
991 */
992void
993lwkt_user_yield(void)
994{
995 thread_t td = curthread;
996 struct lwp *lp = td->td_lwp;
997
998#ifdef SMP
999 /*
1000 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1001 * kernel can prevent other cpus from servicing interrupt threads
1002 * which still require the MP lock (which is a lot of them). This
1003 * has a chaining effect since if the interrupt is blocked, so is
1004 * the event, so normal scheduling will not pick up on the problem.
1005 */
684a93c4
MD
1006 if (mp_lock_contention_mask && td->td_mpcount) {
1007 yield_mplock(td);
3824f392
MD
1008 }
1009#endif
1010
1011 /*
1012 * Another kernel thread wants the cpu
1013 */
1014 if (lwkt_resched_wanted())
1015 lwkt_switch();
1016
1017 /*
1018 * If the user scheduler has asynchronously determined that the current
1019 * process (when running in user mode) needs to lose the cpu then make
1020 * sure we are released.
1021 */
1022 if (user_resched_wanted()) {
1023 if (td->td_release)
1024 td->td_release(td);
1025 }
1026
1027 /*
1028 * If we are released reduce our priority
1029 */
1030 if (td->td_release == NULL) {
1031 if (lwkt_check_resched(td) > 0)
1032 lwkt_switch();
e381e77c
MD
1033 if (lp) {
1034 lp->lwp_proc->p_usched->acquire_curproc(lp);
1035 td->td_release = lwkt_passive_release;
1036 lwkt_setpri_self(TDPRI_USER_NORM);
1037 }
3824f392
MD
1038 }
1039}
1040
1041/*
b9eb1c19
MD
1042 * Return 0 if no runnable threads are pending at the same or higher
1043 * priority as the passed thread.
1044 *
1045 * Return 1 if runnable threads are pending at the same priority.
1046 *
1047 * Return 2 if runnable threads are pending at a higher priority.
1048 */
1049int
1050lwkt_check_resched(thread_t td)
1051{
1052 int pri = td->td_pri & TDPRI_MASK;
1053
1054 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1055 return(2);
1056 if (TAILQ_NEXT(td, td_threadq))
1057 return(1);
1058 return(0);
1059}
1060
1061/*
f1d1c3fa
MD
1062 * Generic schedule. Possibly schedule threads belonging to other cpus and
1063 * deal with threads that might be blocked on a wait queue.
1064 *
0a3f9b47
MD
1065 * We have a little helper inline function which does additional work after
1066 * the thread has been enqueued, including dealing with preemption and
1067 * setting need_lwkt_resched() (which prevents the kernel from returning
1068 * to userland until it has processed higher priority threads).
6330a558
MD
1069 *
1070 * It is possible for this routine to be called after a failed _enqueue
1071 * (due to the target thread migrating, sleeping, or otherwise blocked).
1072 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1073 *
1074 * reschedok is an optimized constant propagated from lwkt_schedule() or
1075 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1076 * reschedule to be requested if the target thread has a higher priority.
1077 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1078 * be 0, prevented undesired reschedules.
8ad65e08 1079 */
0a3f9b47
MD
1080static __inline
1081void
361d01dd 1082_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1083{
b9eb1c19 1084 thread_t otd;
c730be20 1085
6330a558 1086 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1087 if (ntd->td_preemptable && reschedok) {
6330a558 1088 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1089 } else if (reschedok) {
b9eb1c19
MD
1090 otd = curthread;
1091 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1092 need_lwkt_resched();
6330a558 1093 }
0a3f9b47
MD
1094 }
1095}
1096
361d01dd 1097static __inline
8ad65e08 1098void
361d01dd 1099_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1100{
37af14fe
MD
1101 globaldata_t mygd = mycpu;
1102
41a01a4d 1103 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1104 crit_enter_gd(mygd);
9388413d 1105 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1106 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1107 _lwkt_enqueue(td);
1108 } else {
f1d1c3fa 1109 /*
7cd8d145
MD
1110 * If we own the thread, there is no race (since we are in a
1111 * critical section). If we do not own the thread there might
1112 * be a race but the target cpu will deal with it.
f1d1c3fa 1113 */
0f7a3396 1114#ifdef SMP
7cd8d145 1115 if (td->td_gd == mygd) {
9d265729 1116 _lwkt_enqueue(td);
361d01dd 1117 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1118 } else {
e381e77c 1119 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1120 }
0f7a3396 1121#else
7cd8d145 1122 _lwkt_enqueue(td);
361d01dd 1123 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1124#endif
8ad65e08 1125 }
37af14fe 1126 crit_exit_gd(mygd);
8ad65e08
MD
1127}
1128
361d01dd
MD
1129void
1130lwkt_schedule(thread_t td)
1131{
1132 _lwkt_schedule(td, 1);
1133}
1134
1135void
1136lwkt_schedule_noresched(thread_t td)
1137{
1138 _lwkt_schedule(td, 0);
1139}
1140
88ebb169
SW
1141#ifdef SMP
1142
e381e77c
MD
1143/*
1144 * When scheduled remotely if frame != NULL the IPIQ is being
1145 * run via doreti or an interrupt then preemption can be allowed.
1146 *
1147 * To allow preemption we have to drop the critical section so only
1148 * one is present in _lwkt_schedule_post.
1149 */
1150static void
1151lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1152{
1153 thread_t td = curthread;
1154 thread_t ntd = arg;
1155
1156 if (frame && ntd->td_preemptable) {
1157 crit_exit_noyield(td);
1158 _lwkt_schedule(ntd, 1);
1159 crit_enter_quick(td);
1160 } else {
1161 _lwkt_schedule(ntd, 1);
1162 }
1163}
1164
d9eea1a5 1165/*
52eedfb5
MD
1166 * Thread migration using a 'Pull' method. The thread may or may not be
1167 * the current thread. It MUST be descheduled and in a stable state.
1168 * lwkt_giveaway() must be called on the cpu owning the thread.
1169 *
1170 * At any point after lwkt_giveaway() is called, the target cpu may
1171 * 'pull' the thread by calling lwkt_acquire().
1172 *
ae8e83e6
MD
1173 * We have to make sure the thread is not sitting on a per-cpu tsleep
1174 * queue or it will blow up when it moves to another cpu.
1175 *
52eedfb5 1176 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1177 */
a2a5ad0d 1178void
52eedfb5
MD
1179lwkt_giveaway(thread_t td)
1180{
3b4192fb 1181 globaldata_t gd = mycpu;
52eedfb5 1182
3b4192fb
MD
1183 crit_enter_gd(gd);
1184 if (td->td_flags & TDF_TSLEEPQ)
1185 tsleep_remove(td);
1186 KKASSERT(td->td_gd == gd);
1187 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1188 td->td_flags |= TDF_MIGRATING;
1189 crit_exit_gd(gd);
52eedfb5
MD
1190}
1191
1192void
a2a5ad0d
MD
1193lwkt_acquire(thread_t td)
1194{
37af14fe
MD
1195 globaldata_t gd;
1196 globaldata_t mygd;
a2a5ad0d 1197
52eedfb5 1198 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1199 gd = td->td_gd;
37af14fe 1200 mygd = mycpu;
52eedfb5 1201 if (gd != mycpu) {
35238fa5 1202 cpu_lfence();
52eedfb5 1203 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1204 crit_enter_gd(mygd);
df910c23
MD
1205 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1206#ifdef SMP
1207 lwkt_process_ipiq();
1208#endif
52eedfb5 1209 cpu_lfence();
df910c23 1210 }
37af14fe 1211 td->td_gd = mygd;
52eedfb5
MD
1212 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1213 td->td_flags &= ~TDF_MIGRATING;
1214 crit_exit_gd(mygd);
1215 } else {
1216 crit_enter_gd(mygd);
1217 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1218 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1219 crit_exit_gd(mygd);
a2a5ad0d
MD
1220 }
1221}
1222
52eedfb5
MD
1223#endif
1224
8ad65e08 1225/*
f1d1c3fa
MD
1226 * Generic deschedule. Descheduling threads other then your own should be
1227 * done only in carefully controlled circumstances. Descheduling is
1228 * asynchronous.
1229 *
1230 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1231 */
1232void
1233lwkt_deschedule(thread_t td)
1234{
f1d1c3fa 1235 crit_enter();
b8a98473 1236#ifdef SMP
f1d1c3fa
MD
1237 if (td == curthread) {
1238 _lwkt_dequeue(td);
1239 } else {
a72187e9 1240 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1241 _lwkt_dequeue(td);
1242 } else {
b8a98473 1243 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1244 }
1245 }
b8a98473
MD
1246#else
1247 _lwkt_dequeue(td);
1248#endif
f1d1c3fa
MD
1249 crit_exit();
1250}
1251
1252/*
4b5f931b
MD
1253 * Set the target thread's priority. This routine does not automatically
1254 * switch to a higher priority thread, LWKT threads are not designed for
1255 * continuous priority changes. Yield if you want to switch.
1256 *
1257 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1258 * of the td_pri field. The specified priority may also indicate zero or
1259 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1260 *
1261 * Note that we requeue the thread whether it winds up on a different runq
1262 * or not. uio_yield() depends on this and the routine is not normally
1263 * called with the same priority otherwise.
4b5f931b
MD
1264 */
1265void
1266lwkt_setpri(thread_t td, int pri)
1267{
26a0694b 1268 KKASSERT(pri >= 0);
a72187e9 1269 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1270 crit_enter();
1271 if (td->td_flags & TDF_RUNQ) {
1272 _lwkt_dequeue(td);
1273 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1274 _lwkt_enqueue(td);
1275 } else {
1276 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1277 }
1278 crit_exit();
1279}
1280
03bd0a5e
MD
1281/*
1282 * Set the initial priority for a thread prior to it being scheduled for
1283 * the first time. The thread MUST NOT be scheduled before or during
1284 * this call. The thread may be assigned to a cpu other then the current
1285 * cpu.
1286 *
1287 * Typically used after a thread has been created with TDF_STOPPREQ,
1288 * and before the thread is initially scheduled.
1289 */
1290void
1291lwkt_setpri_initial(thread_t td, int pri)
1292{
1293 KKASSERT(pri >= 0);
1294 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1295 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1296}
1297
26a0694b
MD
1298void
1299lwkt_setpri_self(int pri)
1300{
1301 thread_t td = curthread;
1302
4b5f931b
MD
1303 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1304 crit_enter();
1305 if (td->td_flags & TDF_RUNQ) {
1306 _lwkt_dequeue(td);
1307 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1308 _lwkt_enqueue(td);
1309 } else {
1310 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1311 }
1312 crit_exit();
1313}
1314
5d21b981 1315/*
52eedfb5
MD
1316 * Migrate the current thread to the specified cpu.
1317 *
1318 * This is accomplished by descheduling ourselves from the current cpu,
1319 * moving our thread to the tdallq of the target cpu, IPI messaging the
1320 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1321 * races while the thread is being migrated.
ae8e83e6
MD
1322 *
1323 * We must be sure to remove ourselves from the current cpu's tsleepq
1324 * before potentially moving to another queue. The thread can be on
1325 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1326 */
3d28ff59 1327#ifdef SMP
5d21b981 1328static void lwkt_setcpu_remote(void *arg);
3d28ff59 1329#endif
5d21b981
MD
1330
1331void
1332lwkt_setcpu_self(globaldata_t rgd)
1333{
1334#ifdef SMP
1335 thread_t td = curthread;
1336
1337 if (td->td_gd != rgd) {
1338 crit_enter_quick(td);
ae8e83e6 1339 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1340 tsleep_remove(td);
5d21b981
MD
1341 td->td_flags |= TDF_MIGRATING;
1342 lwkt_deschedule_self(td);
52eedfb5 1343 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1344 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1345 lwkt_switch();
1346 /* we are now on the target cpu */
52eedfb5 1347 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1348 crit_exit_quick(td);
1349 }
1350#endif
1351}
1352
ecdefdda
MD
1353void
1354lwkt_migratecpu(int cpuid)
1355{
1356#ifdef SMP
1357 globaldata_t rgd;
1358
1359 rgd = globaldata_find(cpuid);
1360 lwkt_setcpu_self(rgd);
1361#endif
1362}
1363
5d21b981
MD
1364/*
1365 * Remote IPI for cpu migration (called while in a critical section so we
1366 * do not have to enter another one). The thread has already been moved to
1367 * our cpu's allq, but we must wait for the thread to be completely switched
1368 * out on the originating cpu before we schedule it on ours or the stack
1369 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1370 * change to main memory.
1371 *
1372 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1373 * against wakeups. It is best if this interface is used only when there
1374 * are no pending events that might try to schedule the thread.
1375 */
3d28ff59 1376#ifdef SMP
5d21b981
MD
1377static void
1378lwkt_setcpu_remote(void *arg)
1379{
1380 thread_t td = arg;
1381 globaldata_t gd = mycpu;
1382
df910c23
MD
1383 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1384#ifdef SMP
1385 lwkt_process_ipiq();
1386#endif
35238fa5 1387 cpu_lfence();
df910c23 1388 }
5d21b981 1389 td->td_gd = gd;
35238fa5 1390 cpu_sfence();
5d21b981 1391 td->td_flags &= ~TDF_MIGRATING;
9388413d 1392 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1393 _lwkt_enqueue(td);
1394}
3d28ff59 1395#endif
5d21b981 1396
553ea3c8 1397struct lwp *
4b5f931b
MD
1398lwkt_preempted_proc(void)
1399{
73e4f7b9 1400 thread_t td = curthread;
4b5f931b
MD
1401 while (td->td_preempted)
1402 td = td->td_preempted;
553ea3c8 1403 return(td->td_lwp);
4b5f931b
MD
1404}
1405
4b5f931b 1406/*
99df837e
MD
1407 * Create a kernel process/thread/whatever. It shares it's address space
1408 * with proc0 - ie: kernel only.
1409 *
365fa13f
MD
1410 * NOTE! By default new threads are created with the MP lock held. A
1411 * thread which does not require the MP lock should release it by calling
1412 * rel_mplock() at the start of the new thread.
99df837e
MD
1413 */
1414int
1415lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1416 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1417 const char *fmt, ...)
99df837e 1418{
73e4f7b9 1419 thread_t td;
e2565a42 1420 __va_list ap;
99df837e 1421
d3d32139 1422 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1423 tdflags);
a2a5ad0d
MD
1424 if (tdp)
1425 *tdp = td;
709799ea 1426 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1427
1428 /*
1429 * Set up arg0 for 'ps' etc
1430 */
e2565a42 1431 __va_start(ap, fmt);
379210cb 1432 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1433 __va_end(ap);
99df837e
MD
1434
1435 /*
1436 * Schedule the thread to run
1437 */
ef0fdad1
MD
1438 if ((td->td_flags & TDF_STOPREQ) == 0)
1439 lwkt_schedule(td);
1440 else
1441 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1442 return 0;
1443}
1444
1445/*
1446 * Destroy an LWKT thread. Warning! This function is not called when
1447 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1448 * uses a different reaping mechanism.
1449 */
1450void
1451lwkt_exit(void)
1452{
1453 thread_t td = curthread;
c070746a 1454 thread_t std;
8826f33a 1455 globaldata_t gd;
99df837e
MD
1456
1457 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1458 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1459 caps_exit(td);
c070746a
MD
1460
1461 /*
1462 * Get us into a critical section to interlock gd_freetd and loop
1463 * until we can get it freed.
1464 *
1465 * We have to cache the current td in gd_freetd because objcache_put()ing
1466 * it would rip it out from under us while our thread is still active.
1467 */
1468 gd = mycpu;
37af14fe 1469 crit_enter_quick(td);
c070746a
MD
1470 while ((std = gd->gd_freetd) != NULL) {
1471 gd->gd_freetd = NULL;
1472 objcache_put(thread_cache, std);
1473 }
3b4192fb
MD
1474
1475 /*
1476 * Remove thread resources from kernel lists and deschedule us for
1477 * the last time.
1478 */
1479 if (td->td_flags & TDF_TSLEEPQ)
1480 tsleep_remove(td);
79eae878 1481 biosched_done(td);
f8abf63c 1482 dsched_exit_thread(td);
37af14fe 1483 lwkt_deschedule_self(td);
e56e4dea 1484 lwkt_remove_tdallq(td);
c070746a
MD
1485 if (td->td_flags & TDF_ALLOCATED_THREAD)
1486 gd->gd_freetd = td;
99df837e
MD
1487 cpu_thread_exit();
1488}
1489
e56e4dea
MD
1490void
1491lwkt_remove_tdallq(thread_t td)
1492{
1493 KKASSERT(td->td_gd == mycpu);
1494 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1495}
1496
2d93b37a
MD
1497void
1498crit_panic(void)
1499{
1500 thread_t td = curthread;
1501 int lpri = td->td_pri;
1502
1503 td->td_pri = 0;
1504 panic("td_pri is/would-go negative! %p %d", td, lpri);
1505}
1506
d165e668
MD
1507#ifdef SMP
1508
bd8015ca
MD
1509/*
1510 * Called from debugger/panic on cpus which have been stopped. We must still
1511 * process the IPIQ while stopped, even if we were stopped while in a critical
1512 * section (XXX).
1513 *
1514 * If we are dumping also try to process any pending interrupts. This may
1515 * or may not work depending on the state of the cpu at the point it was
1516 * stopped.
1517 */
1518void
1519lwkt_smp_stopped(void)
1520{
1521 globaldata_t gd = mycpu;
1522
1523 crit_enter_gd(gd);
1524 if (dumping) {
1525 lwkt_process_ipiq();
1526 splz();
1527 } else {
1528 lwkt_process_ipiq();
1529 }
1530 crit_exit_gd(gd);
1531}
1532
d165e668 1533#endif